Grounding
Grounding is the practice of answering from real, retrieved information instead of a model's memorized guess.
Category
Industry concept
See all terms
What is grounding?
Grounding means a model produces its answer from actual source material, a document, a database record, a live lookup, rather than from patterns it picked up during training. A model's training data has a cutoff and never includes private company information, so on its own it has nothing current or specific to work from. Grounding fixes that by handing the model a real source at the moment it answers, so the answer traces back to something that actually exists.
How is grounding different from RAG?
Grounding and RAG aren't the same thing: RAG is a pipeline, the retrieve-then-generate steps that fetch matching text and insert it into context window, while grounding is the property those steps are trying to produce, an answer anchored to a real source. A model can be grounded without a RAG pipeline too, a document pasted directly into the conversation grounds the answer just as well. RAG is one way to achieve grounding, not the definition of it.
How does grounding work?
In practice, grounding means retrieving relevant, current information and inserting it into the model's context before it generates a response, whether that's a single file, a set of search results, or a tool output. The model still has to read that material correctly, so grounding lowers the odds of a wrong answer without removing them, a model can still misread or misapply a source it was handed. Grounding is only as good as the source it points to: a grounded answer built on an outdated or wrong document is still wrong, just wrong with more confidence. Qontext keeps that source itself structured, current, and permission-aware in the context repository, so what an agent retrieves is accurate to begin with, not just present.